For a long time, editing an AI image was like working with a painter with no short-term memory. You’d describe a cozy coffee shop and get something lovely. Then you’d ask for one small change, “make the mug red instead of blue,” and they’d start over: new walls, new furniture, a different person holding a red mug.
ChatGPT Images 2.5, released on September 8, 2026, is a big step toward curing that amnesia. It also makes you wait a lot less.
Two versions: Flare and Sunburst
Instead of one model that tries to do everything, OpenAI now offers two.
- Flare is the quick one. It’s the default for most work, and OpenAI says it delivers better images than Images 2.0 in about half the time. For brainstorming, illustrating a newsletter, or trying ideas in a chat, you get results almost as fast as you can type.
- Sunburst is the careful one. It takes longer per image and uses that time to get fine detail, natural lighting, and consistency right. If you’re making product shots, brand assets, or a storyboard where one character has to look the same across ten scenes, Sunburst keeps faces, clothing, and logos stable from edit to edit.
The choice is simple. Use Flare until an image needs to survive many rounds of changes, then switch to Sunburst.
Small features that save real time
Transparent backgrounds. Making an icon or sticker used to mean generating an image on a white background and then erasing the background with another tool. That often left a fuzzy halo around hair or glass. Now you can ask for a transparent background directly and drop the result straight into your slides or website.
@Sketch. Describing a layout in words is surprisingly hard. With the @Sketch tool in ChatGPT, you draw a rough scribble instead: person here, headline there, horizon about a third of the way up. The model follows your layout.
Text that reads correctly. Images 2.0 made readable text in AI images reliable, and 2.5 keeps it. Shop signs, product labels, and app screens come out with properly spelled words, including in Chinese, Japanese, and Korean.
If you came here from Reve
If a link to our old Reve 2.1 review brought you here, here’s what happened. After an investment from OpenAI and the departure of several key researchers to OpenAI, Reve shut down. Its API closed on August 14, 2026, and image generation on reve.com ended on September 27, 2026. Stored images can be downloaded until October 31.
Reve was loved for editing one region without disturbing the rest of the image. Of the tools still running, Images 2.5’s Sunburst model is the closest replacement for that workflow. Krea and Recraft are also worth a look if you prefer an independent tool.
Where it still falls short
- Strict safety filters. OpenAI’s guardrails are conservative. Concepts involving mock violence, dark satire, or horror can be refused where other tools would go ahead.
- Less artistic flair. Photorealism and clean commercial design are its strengths. For painterly, dreamy, atmospheric art, Midjourney V8 still feels more like a real illustrator.
- Precision isn’t free. Long Sunburst editing sessions use up paid plan limits or API budget faster than quick Flare images.
The verdict
For designers, marketers, and anyone who needs images that communicate clearly, with correct text and edits that don’t start from scratch, ChatGPT Images 2.5 remains our #1 image tool. Start with Flare, and switch to Sunburst when the details have to hold.
